Identifying Predictively Essential Information in Machine Learning: An Econometric-Guided Framework with Cross-Dataset Evidence from Housing Markets

Machine-learning performance depends not only on algorithm selection but also on whether the predictor set preserves information essential for reliable prediction. This study develops an econometric-guided cross-dataset predictor-necessity framework that distinguishes predictor importance from predictive necessity through controlled infor-mation removal. Three information environments are compared: a full-information benchmark, an ex ante information-removal specification, and an econometrically screened specification. Random Forest, XGBoost, and artificial neural networks are evalu-ated under common out-of-sample validation conditions across four public housing da-tasets. Econometric screening retained dataset-specific combinations of accessibility, spa-tial location, income, age, quality, living space, condition, and related structural attributes. Controlled removal identified accessibility information in UCI Real Estate Valuation, spa-tial-location information in California Housing, and property-quality and living-space information in Ames Housing and King County as consequential valuation channels. Across twelve dataset-algorithm combinations, the econometrically screened specification retained an average of 98.93% of full-information predictive R², whereas the controlled in-formation-removal specification retained 92.74%. Predictive deterioration therefore de-pended more on the substantive information removed than on the number of predictors excluded. The findings demonstrate that essential predictors are market-specific, whereas the procedure for identifying consequential information is transferable across datasets and learning architectures.